Aircraft engine compartment small target flame detection method and system and medium
By combining ViBe and the improved YOLOv5 algorithm, the false alarm rate problem in aircraft engine compartment fire detection was solved, achieving efficient and accurate detection of small target flames, reducing the false alarm rate, and ensuring the visual confirmation of engine compartment fires.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-26
- Publication Date
- 2026-03-10
AI Technical Summary
In existing technologies, fire detection in aircraft engine compartments is mainly carried out through aerodynamic lines, which has a high false alarm rate. Furthermore, since the engine compartment is a power system, stopping or rapidly landing can result in extremely high losses.
The ViBe algorithm is combined with the improved Yolov5 algorithm. The ViBe algorithm is used to detect moving foregrounds and suppress interference from stationary fire-like objects. The improved Yolov5 algorithm adds a CBS module, a multi-scale FPN-PAN structure and an SA attention mechanism to improve the detection capability of small targets.
It reduced the occurrence of false alarms, improved the detection accuracy and response speed of small target flames, reduced the false alarm rate, and ensured the effective detection of engine compartment fires.
Smart Images
Figure CN121640243A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of fire detection, and relates to an aircraft engine compartment small target flame detection method, system and medium, in particular to an improved small target flame detection method, system and medium based on a combination of a ViBe algorithm and a Yolov5 algorithm. BACKGROUND
[0002] The engine compartment has a complex internal structure, various finished accessories, pipelines and cables are distributed therein, and the flame picture is divided into many small flame targets by the partition. At the same time, there are a large amount of flammable materials such as fuel and electrical lines in the compartment, and complex environments such as high temperature and high pressure are generated during operation, and the occurrence of fire has the characteristics of suddenness and rapid spread. Therefore, it is crucial to effectively detect the fire in the aircraft engine compartment.
[0003] At present, the aircraft engine compartment fire detection is mainly performed through a pneumatic line, but it has a high false alarm rate. Since the engine compartment is a power system, stopping work or rapid landing will cause high loss, and therefore it is necessary to research a fire detection method based on machine vision to visually confirm the engine compartment fire response. SUMMARY
[0004] The present application aims to provide an aircraft engine compartment small target flame detection method, system and medium, which is used for the aircraft engine compartment fire detection task, and solves the problem of high false alarm rate in the current aircraft engine compartment fire detection mainly through a pneumatic line.
[0005] The technical scheme of the present application is as follows: An aircraft engine compartment small target flame detection method comprises the following steps: A ViBe algorithm model is used to detect an engine compartment fire image with a moving foreground; A Yolov5 improved algorithm model is used to identify the flame of the engine compartment fire image detected by the ViBe algorithm model.
[0006] Further, the Yolov5 improved algorithm model comprises a backbone network, a neck structure and a detection head.
[0007] Further, the engine compartment fire image is input into the ViBe algorithm model for detection, the image with the moving foreground is sent into the ViBe algorithm model, the backbone network in the ViBe algorithm model extracts features from the input image; the neck structure fuses the features extracted by the backbone network to enhance the representation ability of the features and provide better feature input for the detection head; the detection head performs target positioning and classification based on the feature map and outputs the final detection result.
[0008] Further, the Yolov5 improved algorithm model is used for flame recognition on the current image of the engine compartment fire, if the image contains flame, a fire alarm is sent, otherwise, the next frame of image is continuously detected.
[0009] Further, the backbone network comprises a CBS module and a C3 module, wherein the CBS module comprises a Conv convolutional layer, a BN batch normalization layer and a SiLU activation function layer, the CBS module is added in the backbone network, the CBS module extracts features in the input data through the Conv convolutional layer, the BN batch normalization layer normalizes the features to ensure the stability of the data distribution, and the SiLU activation function layer enhances the nonlinear expression ability of the model to improve the feature extraction capability of the Yolov5 improved algorithm model.
[0010] Further, the multi-scale FPN-PAN structure is used in the neck structure, the up-sampling feature map and the down-sampling feature map are added in the multi-scale FPN-PAN structure, and the multi-scale FPN-PAN structure is expanded through fusion processing.
[0011] Further, the target detection branch is added in the detection head, and the SA attention mechanism is added in the detection branch.
[0012] Further, the class loss function and the target loss function in the detection head are both selected as the BCE loss function, and the position loss function is selected as the SIoU loss function.
[0013] A system applied to the small target flame detection method of the aircraft engine compartment, comprising: An image motion foreground detection module: an engine compartment fire image with motion foreground is detected through a ViBe algorithm model; A flame recognition module: the engine compartment fire image detected through the ViBe algorithm model is subjected to flame recognition through a Yolov5 improved algorithm model.
[0014] A computer storage medium, the storage medium stores a computer program, and the computer program is used for executing the small target flame detection method of the aircraft engine compartment.
[0015] Advantages At present, due to the shielding problem, the flame target is small, and in order to meet the detection requirements of the initial stage of the flame, the application provides a small target flame detection method, system and medium for an aircraft engine compartment, which is specifically a small target flame detection method and system based on the combination of a ViBe algorithm and an improved Yolov5 algorithm. The application combines the ViBe algorithm with the improved Yolov5 deep learning algorithm, adopts the ViBe algorithm to perform motion foreground detection, suppresses the interference of static fire-like objects, and reduces the occurrence of false positives. The Yolov5 deep learning algorithm is improved, so that it is more suitable for small target detection. A CBS module is added in the backbone network to improve the feature extraction capability of the model; a multi-scale FPN-PAN structure is adopted in the neck structure to improve the capture capability of the model for very small targets; on the basis of the three detection branches of the Yolov5s model, a very small target detection branch is added to improve the focusing capability of the model for small targets. The class loss function and the target loss function both select the BCE loss function. The position loss function selects the SIoU loss function. BRIEF DESCRIPTION OF DRAWINGS
[0016] Figure 1 is a small target flame detection algorithm structure diagram of the application; Figure 2 is a CBS module structure diagram in the application; Figure 3 is a module processing flowchart in the backbone network in the application; Figure 4 is a multi-scale FPN+PAN structure diagram in the application. DETAILED DESCRIPTION
[0017] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the application.
[0018] One embodiment of the application provides a small target flame detection method based on the combination of a ViBe algorithm and an improved Yolov5 algorithm, which includes the following steps: The engine compartment fire image is input into the ViBe algorithm model for detection, and the image with the determined motion foreground is sent to the deep learning-based fire detection network for flame recognition. The deep learning-based fire detection network is an improved Yolov5 algorithm model.
[0019] The ViBe algorithm model is used to detect the motion foreground to suppress the interference of static fire-like objects and reduce the occurrence of false positives.
[0020] The deep learning algorithm adopts a small target flame detection algorithm model improved based on Yolov5.
[0021] The small target flame detection algorithm model improved based on Yolov5 includes a backbone network, a neck structure, and a detection head. The backbone network, the neck structure, and the detection head are improved respectively, and the position loss function of the detection head is modified.
[0022] The backbone network includes a CBS module and a C3 module, wherein the CBS module includes a Conv convolutional layer, a BN batch normalization layer, and a SiLU activation function layer. The CBS module is added in the backbone network. The CBS module extracts features in the input data through the Conv convolutional layer, the BN batch normalization layer normalizes the features to ensure the stability of the data distribution, and the SiLU activation function layer enhances the nonlinear expression ability of the model, thereby improving the feature extraction capability of the model.
[0023] In the neck structure, a multi-scale FPN-PAN structure is adopted, an up-sampling feature map and a down-sampling feature map are added on the multi-scale FPN-PAN structure, and fusion processing is performed for expansion.
[0024] In the detection head, a target detection branch is added, and an SA attention mechanism is added in the detection branch.
[0025] The class loss function and the target loss function in the detection head both select the BCE loss function, and the position loss function selects the SIoU loss function.
[0026] In this embodiment, the ViBe algorithm model is used to detect the moving foreground to suppress the interference of the static fire-like objects and reduce the occurrence of false positives.
[0027] In this embodiment, the deep learning algorithm adopts a small target flame detection algorithm improved based on Yolov5. The small target flame detection algorithm improved based on Yolov5 improves the backbone network, the neck structure, and the detection head respectively, and modifies the loss function.
[0028] In this embodiment, a CBS module is added in the backbone network. The CBS module extracts features in the input data through the Conv convolutional layer, the BN batch normalization layer normalizes the features to ensure the stability of the data distribution, and the SiLU activation function layer enhances the nonlinear expression ability of the model, thereby improving the feature extraction capability of the model.
[0029] In this embodiment, a multi-scale FPN-PAN structure is adopted in the neck structure. An up-sampling feature map is added on the original Yolov5s neck structure, and fusion processing is performed for expansion.
[0030] In the embodiment, a target detection branch is added in the detection head, and a SA attention mechanism is added in the detection branch.
[0031] In the embodiment, the category loss function and the target loss function are both selected as the BCE loss function, and the position loss function is selected as the SIoU loss function.
[0032] The second embodiment of the application comprises Figure 1 As shown in the figure, a small target flame detection algorithm based on the combination of the ViBe algorithm and the improved Yolov5 algorithm is provided. The specific implementation manner is as follows: a small target flame detection algorithm based on the combination of the ViBe algorithm and the improved Yolov5 algorithm inputs the engine compartment fire image into a traditional detection algorithm for detection, and the image with the determined moving foreground is sent into a deep learning-based fire detection network. The backbone network extracts features from the input image. The neck structure fuses the features extracted by the backbone network, enhances the representation ability of the features, and provides better feature input for the detection head. The detection head performs target positioning and classification based on the feature map, and outputs the final detection result. The deep learning-based fire detection network is used for flame recognition. If the flame is contained, a fire alarm is issued; otherwise, the next frame of image is continuously detected.
[0033] In the embodiment, the ViBe algorithm is used to detect the moving foreground to suppress the interference of the static fire-like objects and reduce the occurrence of false positives.
[0034] In the embodiment, the deep learning algorithm adopts the small target flame detection algorithm based on the improved Yolov5.
[0035] In the embodiment, the small target flame detection algorithm based on the improved Yolov5 improves the backbone network, the neck structure and the detection head, and modifies the position loss function.
[0036] The CBS module is added in the backbone network. The CBS module extracts the features in the input data through the Conv convolution layer, normalizes the features through the BN batch normalization layer, ensures the stability of the data distribution, and enhances the nonlinear expression ability of the model through the SiLU activation function layer, thereby improving the feature extraction ability of the model. In the CBS module, the convolution kernel is 3x3 with a step of 2, the batch normalization and SiLU activation function are used for shallow feature extraction, and the P1 feature map with a size of 160x160x64 is obtained. Then, the C3 module is used for deep feature extraction. As shown in the figure, Figure 3As shown, P1 enters two branches in C3, and after CBS processing, P1_0 and P1_1 feature maps are output, and the bottleneck structure outputs a feature map with a size of 160x160x64. P1_1 is a bottleneck structure processed by two CBSs and residual edges containing its own information, and a P1-2 feature map with a size of 160x160x64 is obtained. P1_0 and P1_2 are spliced in the channel size (160x160x128), and then CBS processing is performed again to obtain a P2 feature map with a size of 160x160x64. The backbone of this model contains 4 such CBSs and C3s in series, which respectively generate P2, P3, P4, and P5 feature maps.
[0037] A multi-scale FPN-PAN structure is used in the neck structure, as shown in Figure 4 The FPN passes down semantic information, and the PAN passes up and combines location information. After convolution processing of P5, P5_md is obtained, and P5_md is up-sampled and spliced with P4 in the channel dimension to obtain P4_md. Similarly, P3-md and P2_md can be obtained. After convolution processing of P2_md, P2_out is obtained, and P2_out is down-sampled and spliced with P3_md in the channel dimension to obtain P3_out, and this cycle is repeated to obtain P4_out and P5_out. Through the fusion processing of the expanded P2 feature map, the output P2_out, P3_out, P4_out, and P5_out feature maps of different scales contain strong semantic information and location information, which improves the model's ability to capture very small targets.
[0038] A target detection branch is added in the detection head, and an SA attention mechanism is added in the detection branch. As shown in Figure 1 P2_out, P3_out, P4_out, and P5_out are processed in the head to obtain four feature maps of different scales, H2, H3, H4, and H5. Through 4 times, 8 times, 16 times, and 32 times of down-sampling of the input image, feature maps with sizes of 160x160, 80x80, 40x40, and 20x20 are output, corresponding to 4x4, 8x8, 16x16, and 32x32 pixels of the original image, respectively, which expands the detection range of the original Yolov5s. By adding an SA attention mechanism in the small target and small target detection branch, the focusing ability of the model on small targets is improved.
[0039] Both the category loss function and the target loss function select the BCE loss function, and the position loss function selects the SIoU loss function, which improves the convergence performance of the regression loss and reduces the final regression loss error of the target frame, and judges the feature map to obtain the prediction result.
[0040] The third embodiment of the present application provides a system applied to the small target flame detection method of the aircraft engine compartment, comprising: An image motion foreground detection module: an engine compartment fire image with a motion foreground is detected through a ViBe algorithm model; A flame recognition module: the engine compartment fire image detected through the ViBe algorithm model is subjected to flame recognition through a Yolov5 improved algorithm model.
[0041] The fourth embodiment of the present application provides a computer storage medium, wherein the storage medium stores a computer program, and the computer program is used for executing the small target flame detection method of the aircraft engine compartment.
[0042] In summary, the present application combines the ViBe algorithm with the Yolov5 improved deep learning algorithm, adopts the ViBe algorithm to perform motion foreground detection, suppresses the interference of static fire-like objects, and reduces the occurrence of false positives. The Yolov5 deep learning algorithm is improved to make it more suitable for small target detection.
[0043] The above is only a specific embodiment of the present application, and the present application is described in detail. The part not described in detail is a conventional technology. However, the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. The protection scope of the present application should be subject to the protection scope of the claims.
Claims
1. A method for detecting flames on small targets in an aircraft engine compartment, characterized in that, The method comprises the following steps: The ViBe algorithm model is used to detect the engine cabin fire image with motion foreground; The Yolov5 improved algorithm model is used to identify the flame of the engine cabin fire image detected by the ViBe algorithm model.
2. The aircraft engine nacelle small target flame detection method of claim 1, wherein, The Yolov5 improved algorithm model comprises a backbone network, a neck structure and a detection head.
3. The method of claim 2, wherein, The engine cabin fire image is input into the ViBe algorithm model for detection, and the image with motion foreground is sent into the ViBe algorithm model. The neck structure fuses the features extracted by the backbone network to enhance the representation ability of the features and provide better feature input for the detection head. The detection head performs target positioning and classification based on the feature map and outputs the final detection result.
4. The method of claim 2, wherein, After the Yolov5 improved algorithm model identifies the flame of the current engine cabin fire image, if the image contains flame, a fire alarm is sent out; otherwise, the next frame of image is detected.
5. The method of claim 2, wherein, The backbone network comprises a CBS module and a C3 module, wherein the CBS module comprises a Conv convolutional layer, a BN batch normalization layer and a SiLU activation function layer.
6. The method of claim 5, wherein, In the backbone network, the CBS module is added, the CBS module extracts the features in the input data through the Conv convolutional layer, the BN batch normalization layer normalizes the features to ensure the stability of the data distribution, and the SiLU activation function layer enhances the nonlinear expression ability of the model to improve the feature extraction ability of the Yolov5 improved algorithm model.
7. The method of claim 6, wherein, In the neck structure, a multi-scale FPN-PAN structure is used, and up-sampling feature maps and down-sampling feature maps are added to the multi-scale FPN-PAN structure and fused for expansion.
8. The method of claim 7, wherein, In the detection head, a target detection branch is added, and a SA attention mechanism is added to the detection branch.
9. A system for use in the method of claim 1-8, wherein, In the detection head, the class loss function and the target loss function are both selected as the BCE loss function, and the position loss function is selected as the SIoU loss function. The method comprises the following steps: The image motion foreground detection module detects the engine cabin fire image with motion foreground through the ViBe algorithm model; 10. A computer storage medium, characterized in that, The flame identification module identifies the flame of the engine cabin fire image detected by the ViBe algorithm model through the Yolov5 improved algorithm model. The storage medium stores a computer program, and the computer program is used to execute the aircraft engine cabin small target flame detection method in any one of claims 1-8.